Distributional Reinforcement Learning with Regularized Wasserstein Loss Ke Sun

Empirically, we show that SinkhornDRL consistently outperforms or matches existing algorithms on the Atari games suite and particularly stands out in the multi-dimensional reward setting.